Anecdotally, it is often said that the initial development of AI in corporate treasury and finance is led by a treasury tech enthusiast seeking to move faster and get their hands dirty, unsatisfied with first party tools and wanting something unique and specific.

At energy company Shell, that person is treasury manager and tech savvy Christopher Lim, who is behind the development of an internal AI agent that answers both run-of-the-mill and more complex finance and treasury questions, ranging from policy and guidance to how to issue bank guarantees.

The AI has access to procedural documents, supports knowledge retention, serves as a teaching source and provides advice to internal stakeholders, freeing up time for treasury staff to focus on more strategic elements of the job.

“When someone new joins and they are not sure about something, they can ask the AI to ease the learning curve,” says Lim who joined Shell 18 years ago and has developed the agent alongside his day job in a process that he says has required technical knowledge like programming logic, understanding data security and licencing, and a deep business knowledge.

“You have to know what you want the AI to do,” he explains.

The technology has gone through various stages of evolution since Lim received the green light to develop something bespoke using in-house LLM technology a few years ago. In its first iteration, the agent was limited in a highly constrained “box” for security reasons. As a result, it was prone to hallucinations, pulling information from incorrect sources and frequent “happy failures.” Over time, usage of the LLM declined as people stopped relying on it.

This early setback led to a key insight: accessibility of the user interface was critical to success. Lim redesigned the solution, so that the AI could be directly accessed through Microsoft Teams, significantly improving ease of use and adoption. At the same time, its ability to accurately answer questions was scaled by building a broader and more robust knowledge base in collaboration with different teams across the treasury function.

Today, Lim reflects that by diving in early, Shell is better positioned to be at the forefront of integrating AI. Looking ahead he expects the agent to evolve beyond providing static information such as policies and procedures, instead connecting to databases to answer questions dynamically. Potential future use-cases include advising on cash positions in-country, bank relationships and issued guarantees in a specific geography to assist treasury staff do their job in real time.

He also believes that, with the rapid progress towards more advanced forms of AI, the technology will increasingly be able to reason. Rather than simply reporting cash amounts, it will suggest strategies in response to an expiring bank guarantee or the pros and cons of splitting a cash deposits across accounts to maximise the return. He also predicts that, in the near future, the technology will ultimately be able to execute transactions and communicate with banks through systems such as TMS.

“If the recent progress made in the AI space is any indication,” he says, “this is closer to happening than I first thought.”

Data first

The list of things to get right for adventurous experimenters like Lim is long and detailed.

Rather than try to boil the ocean, companies should start small and pick one painful, high-volume process like bank reconciliation, accounts receivable forecasting or FX aggregation, to get the AI ball rolling. They should then focus on data readiness within that segment, building the data pipeline before building the model. At Shell, the challenge was finding a balance between data security (one early manifestation involved logging into a secure website – a process that put people off) and technical advancement.

Gathering data from ERPs and multiple bank relationships requires real heavy lifting, even with APIs. Moreover, even once the data is connected, it still has to be standardised or normalised because ERP structures are different and every bank uses different naming conventions.

“Data readiness is the #1 barrier to AI integration. It takes between 12-24 months of standardised, historical data before you can start seeing some good use of AI,” Nita Baindur tells Treasury Today Group. Prior to retiring last year as Associate VP, Asst. Treasurer at Agilent Technologies, she consulted closely with treasury vendors and banking partners to assess their AI strategies and discuss emerging use cases.

For all this, she still advises against stalling progress in a quest for perfect data. “The technology is smart and learns over time. It is only necessary to clean the initial data. AI can actually help identify and fix data quality issues over time.”

Teaching staff

Teaching staff how to use the technology is another essential element to success. Integration involves teaching the right skills in a process that goes beyond just typing in a few questions – the AI won’t evolve if people, and processes, don’t evolve alongside it – but this only comes if staff are reassured the technology won’t lead to job displacement.

It has taken deliberate effort to encourage even tech-savvy treasury professionals in Microsoft’s Seattle treasury operation to integrate AI into their daily work, says Kathy Brustad, Director, Treasury, who leads AI for treasury. Brustad notes that adoption is far from linear. While a small group quickly embraces AI and reshapes their workflows within weeks, broader adoption typically follows as others see tangible benefits from their peers. For many, shifting to an AI-first approach requires sustained reinforcement, practical examples and time. And she adds that smaller groups typically move more gradually.

“As people share real outcomes and continue experimenting, momentum builds,” she says. “And as the models improve, the value becomes more evident, making it easier to drive broader, sustained adoption.”

As people share real outcomes and continue experimenting, momentum builds, and as the models improve, the value becomes more evident, making it easier to drive broader, sustained adoption.

Kathy Brustad, Director, Treasury, Microsoft

At Microsoft, treasury staff use AI to help with emails, meetings and writing documents; it also supports cash flow forecasts, hedging strategies and credit collection, a process-heavy, labour-intensive function. “In a human-led and agent-operated process, agents now perform all the manual part of a credit check for us like collecting data on customer payment history and writing up the report. We are careful to make sure the agent explains why it came to its recommendation as part of responsible AI practice, and then the human makes the final decision,” says Brustad.

Most recently, the treasury team has developed an Adam Smith Award-winning tool that scours global news feeds to support quick and agile decision making, surfacing global news that might be high risk and impact cash collection, banking relationships or portfolio management. Of course, this level of confidence sits easier in a tech firm than companies with less IT expertise, but Brustad also believes the technology is within reach of everyone. It is just a question of getting familiar with what is available – and getting started using it.

Working with partners

Working closely with partners is another critical component to success. Small companies already tend to buy off the shelf treasury SAAS tools that eliminate the manual grind of bank reconciliations, payment approvals and forecasting. In larger companies, automation is already in place thanks to the suite of treasury fintech providers. As TMS providers continue to develop agentic AI, they will roll it out to their customers, enabling treasury teams to tap into the technology much faster than if they build something in-house.

At large corporates, Baindur observes that AI is now being layered onto existing systems to give real time visibility to cash positions, optimise liquidity and recommend hedge ratios while on the accounts receivable/payable side, it is protecting against anomalies and fraud. The next challenge involves integrating multiple ERPs into the process to bring complex bank and currencies relationships into the orbit of this high-speed risk optimisation.

She says it’s also possible to have a hybrid solution that draws on internal and external expertise. For example, the accounts receivables team at Agilent (which sat in a shared service centre rather than treasury) developed its own internal AI model, opting not to buy a product from its banking partner. At the company she was involved in reviewing vendor roadmaps, looking at product demos and discussing the most important AI use cases with providers like AtlasFX and 360T, FIS and Kyriba (FireApps). Today she also observes that larger corporates are sometimes rolling out AI using their own internally developed tools alongside technology developed by their banking partners and treasury system vendors.

Building trust – why explainability is essential

Progress also depends on trust which, in turn, relies on explainability behind the answers that AI provides. It means the human team must always be able to track in the backend how AI came to its conclusions so that everyone understands “from start to finish” how it calculated its exposures. “If AI predicts something, you must be able to explain how it came up with that forecast. If you can’t, you will end up going back to spreadsheets,” says Baindur.

She adds that another essential pillar to building trust comes from measuring the business outcome from an AI decision. This could be improved forecast accuracy or lower hedge costs, or the best bank to book a trade with based off historical quotes in a particular currency and over a particular timeframe.

Rather than a fancy new dashboard, it is these kinds of hard numbers that provide the buy-in from internal audit, SOX (Sarbanes-Oxley) and FP&A teams, ensuring that AI accurately integrates rules and policies, that ultimately leads to CFO and CEOs endorsement and budget. “It is evidence that drives AI investment and makes it an enterprise priority,” she says.

Explainability includes looking critically at outputs, a process AI aficionados say requires both institutional knowledge and judgement, gleaned over a career in treasury. In another argument for human judgement, one data point the technology still struggles to read is the nuance between casual and formal communication.

Execution

It’s one of the reasons AI is not carrying out trade execution yet, even in the most sophisticated treasury teams.

Brustad says cash flow forecasting – where AI is already surfacing insights such as optimising cash balances or recommending fund disbursements to subsidiaries – is a natural entry point for deeper AI execution. But she stresses that moving from insight to autonomous action requires careful progression.

While Microsoft has used machine learning in cash forecasting for years, the focus is now shifting to the next stage of human–AI collaboration: from AI-assisted work, to human-led agents, and ultimately to human-led, agent‑operated processes, where agents can execute end-to‑end workflows within defined boundaries.

“In some areas, like lower-risk fund disbursements for daily operations, there is real potential to move toward agent-driven execution,” she says. “But the challenge is defining the right guardrails – where to set boundaries, where to introduce kill switches, and how these decisions play out beyond treasury in a broader enterprise context.”

For now (or at least at the time of writing) she concludes that Microsoft is in ideation and art of the possible stage on this. “Today, humans still execute, with agents supporting decision-making,” she notes. “Moving to auto-execution is an exciting possibility, but it requires rigorous testing and strong controls. As we get the guardrails right – and as the technology continues to improve – it becomes much easier to responsibly move towards human led, agent operated processes.”